They argued about page 32
For two days the internet laughed at a line about time travel in Anthropic's IPO filing, a line nobody who has actually seen the filing has reported. The lines that have been reported are less funny and more important. They say the most valuable safety lab in the world is, underneath everything, a tenant, and on the same evening its biggest new landlord told Wall Street it now makes more money renting out intelligence than flying rockets.
On Tuesday night a screenshot started going round. It was presented as page 32 of Anthropic's S-1, the long prospectus a company has to put in front of the SEC before it can sell shares to the public, and it carried a sentence that read like it had wandered in from a paperback: "Sufficiently advanced systems could discover methods of transmitting information, matter or autonomous systems backward in time, potentially allowing future AI systems to influence events preceding their own creation." By Wednesday it had become a trending story on X, with the jokes you would expect about Roko's Basilisk and the Terminator, a few people asking where they could buy time-travel insurance for their IPO risk model, and a smaller number of people asking the obvious question, which is whether it was real.
I went looking, and I can't stand it up. The filing isn't public. Nothing is on EDGAR, the SEC's public database, and the only people who have described the draft are the reporters who have been shown it, at Reuters in September and at the Wall Street Journal this week. None of them mention time travel. Reuters wrote at length about the risk section when it first saw the draft: about 80 of its 261 pages, nearly double the 48 pages spent on the business, with warnings that models could "resist shutdown", "conceal or manipulate information" and produce behaviour "resembling blackmail". That is strange enough for a document meant to sell shares, and Reuters quoted it carefully. A line about sending matter into the past would have been the first thing anyone quoted. Grok, answering people under the original posts, kept saying the same thing: no public copy, no report from anyone who has read it that matches the screenshot. Maybe it's in there. I'd bet against it, and either way nobody passing it around has shown where it came from.
I don't want to be a scold about a joke. The reason it caught is that it felt plausible, and it felt plausible because Anthropic really has filed the most frightening risk section any company has ever put in front of investors, and people have stopped being able to tell where its real worries end. That's interesting in itself. But the cost of the joke is that the real document, the parts of it we do know about, got far less attention than a sentence that probably doesn't exist. So here are the parts we know about.
The numbers that were reported
The Journal's Anissa Gardizy saw a new draft and published on Tuesday. Revenue in the first half of 2026 was $16.3 billion, up from $1.1 billion in the same six months a year earlier. That is not a typo, and it is the kind of growth curve people use the word "unprecedented" about and then feel slightly embarrassed because it really is. The net loss for the half was $80.7 billion, which is the number that made the headlines, and which is mostly not a loss in the way most people would understand one. About $72 billion of it is a non-cash accounting charge, the paper cost of convertible notes and other instruments Anthropic sold to raise money becoming more valuable as the company did. The business that grew fifteenfold also got more expensive to have lent money to, and the accounts have to record that.
The numbers I'd look at instead are smaller and quieter. Roughly 20 per cent of that $16.3 billion came from two customers, unnamed, and the draft says a small group of customers will keep making up a large share of revenue for the foreseeable future. Anthropic spent nearly $9 billion on compute and infrastructure in those six months, more than the $7.3 billion it spent in the whole of 2025. It has about $450 billion of cloud compute contracts over the next decade, which the Journal broke down as $111 billion with Google Cloud, $110 billion with Amazon Web Services and $31 billion with Microsoft, and it has signalled it could spend up to $750 billion on compute in the coming years once deals still being negotiated are counted. The filing also says, according to the Journal, that damage from a severe AI security incident could exceed what its insurance covers.
And then there is the sentence I think matters most, the one I'd have put on page 32 if I were writing the screenshot. The draft acknowledges that Anthropic has asked for more computing capacity than the big cloud companies were willing to give it, and that this is why it has started turning to smaller cloud providers, and building some of its own.
What a tenant looks like
Put those together and you get a picture that doesn't quite match how we talk about the labs. We talk about Anthropic as one of three or four companies deciding the shape of the future, and in the narrow sense of who trains the best models that may well be true. But a company whose largest commitments run to its competitors' clouds, which needs more capacity than those clouds will sell it, and which leans on two customers for a fifth of its money, is in an important sense a tenant. It is a brilliant, fast-growing, extremely well-funded tenant, and I don't say that to belittle it. But tenants don't set the rent.
There's another number Reuters pulled out of the filing that fills this in. In 2025, 47 per cent of Anthropic's revenue came through its big tech partners, mainly Amazon and Google, with Broadcom and Microsoft. Those partners aren't just where Anthropic gets its computing power. They're also the shop window it sells from. The same two clouds supply most of the power and most of the sales channel, and Google, at least, also builds a model that competes with Claude head on.
None of this is a secret, exactly. People in the industry have known for a while that the hyperscalers hold the cards. What the S-1 does is make it a disclosure, a thing Anthropic's lawyers decided investors had a right to know, with dollar figures attached. A company that describes catastrophic and existential risk in the same document is also telling you, in plainer language, that it doesn't fully control the one input it can't do without.
The landlord reported the same evening
The timing was almost too neat. A few hours after the Journal story, SpaceX, which bought Elon Musk's xAI earlier this year and now runs it as a segment called Super Intelligence, reported its third quarter. Revenue was $14.6 billion, up 161 per cent. The Super Intelligence segment brought in $7.56 billion, up 830 per cent on a year earlier and up 195 per cent on the previous quarter, which made it bigger than Starlink's connectivity business at $5.25 billion. For the first time the AI arm made an operating profit, $431 million, and it would have been more without $1.2 billion of one-off stock compensation from the $60 billion Cursor acquisition.
Where did that money come from? SpaceX's own release puts "the activation of new cloud services agreements" first, ahead of Cursor and Grok. CNBC reported that the company has been renting out capacity to Google, Anthropic and others. TechCrunch's reporting put the size of what SpaceX now runs at 1.7 gigawatts of compute. So one of the firms on the other side of Anthropic's "we couldn't get enough from the hyperscalers" sentence is, by these accounts, a rocket company with a data centre outside Memphis.
Musk said on the call that he spends four days a week at the Colossus II site, and he said the thing that a lot of people in this industry have been saying more quietly for a year: "the constraint on deployment of super intelligence is more in power and cooling and infrastructure than it is in chips." SpaceX spent $19.6 billion on capital projects in the quarter and its finance chief said the fourth quarter would be $35 to $40 billion. Musk also predicted annual recurring revenue of between $500 billion and $1 trillion by the end of 2027, which is the sort of range you should hold loosely, because it is a guess spanning a factor of two from a man whose guesses have a long history. But the direction is not a guess. The company that used to be the AI lab with the most to prove has become, in about two quarters, one of the places other AI labs go to rent their thinking.
And the next morning, a lab walked into its rival's shop
On Wednesday morning OpenAI said its frontier models and Codex will be sold through Google Cloud's Model Garden within weeks, paid for out of customers' existing Google cloud commitments, and that OpenAI models would come to Google's Antigravity coding tool and to Gemini Agent as well. Microsoft, OpenAI's blog post noted in its last line, "remains OpenAI's primary cloud partner". Reuters pointed out that OpenAI's exclusive distribution deal with Microsoft ended in April, that it had already gone to AWS, and that Google was the last of the big three.
On the surface that's a story about choice, and Thomas Kurian said so: give customers lots of models and let them pick. Underneath, it's the same story as the S-1. The labs make the most celebrated product in the world, and they increasingly reach their customers through the cloud companies' storefronts, on the cloud companies' terms, using money the customers have already promised to the cloud companies. If you are a large bank, you don't buy Claude or GPT so much as you spend down your Google commitment on one or the other.
Over in Berlin the same day, Jensen Huang used his keynote at Nvidia's GTC conference to say Europe has about 5 per cent of the world's AI computing capacity against about 75 per cent in the United States, and launched a container-sized data-centre module to help close that gap. It was a reminder that the tenant problem isn't only a company problem. Whole continents are renting too.
What I take from it
I don't think the right reaction to any of this is alarm about Anthropic in particular. It's growing faster than almost any company ever has, and the honesty of its filing is, if anything, a point in its favour. Most companies hide the things that could sink them in boilerplate. This one put shutdown resistance in a prospectus.
What I'd take away is a correction to the story we usually tell. We've spent three years talking about AI power as if it sits with whoever has the best model, which is why every launch gets covered like a coronation. The documents this week, the real ones, keep pointing somewhere else. They point to whoever owns the turbines and the cooling and the contracts, to the companies that sell the electricity-shaped thing every lab needs and that run the shops the labs sell through. Some of those companies also build models, and one of them is now a rocket company. If you want to know who is deciding the shape of this, the clearer answer this week is the landlords.
That makes the time-travel screenshot almost a fitting joke. A line about AI reaching back to change the past was easier to believe, and much more fun to share, than a line about Anthropic not being able to buy enough computers. The second one is the one that's actually in the filing, and it says more about where things are heading.